Automating Statistical Inference for Modern Probabilistic Models

How do you reconstruct an image of a black hole using only noisy telescope measurements in the Fourier domain? How do you predict the cure state of a carbon fiber aircraft wing as it cures in an autoclave using a few faulty thermocouples? In theory, Bayesian probabilistic models are the tool for the job: they can capture the complex latent relationships, realistic data generating mechanisms, and challenging uncertainty structure exhibited by these modern inference problems. But in practice, computational methods for Bayesian inference often fail silently on these sorts of problems, even after significant expert tuning. In this talk, I’ll present a few recent advances in Bayesian computational inference from my group—variational nonreversible parallel tempering and autostep involutive Markov chain methods—that provide reliable and efficient inferential results with little to no user input, and are already having an impact in real scientific problems.